Google DeepMind researcher says LLMs can't jump to new science
Tom Zahavy argues language models can derive and optimize, but lack the embodied leap needed to invent new scientific axioms.
By Wei-Lin Zhao 1 week agoSection
AI
Cisco Provenance Explorer fingerprints lineage for nearly 900 open models
Cisco released a free database to check open-model ancestry, licenses and scan counts as enterprises weigh self-reported AI metada...
By Renata Fuchs 1 week ago
Microsoft AI specialist models put cost ahead of frontier chasing
Mustafa Suleyman says Microsoft AI is training smaller field-specific models, using orchestrators to reserve OpenAI systems for ha...
By Colin Brandt 1 week agoSmartcat Content Update Coworker launches for multilingual training updates
Smartcat’s new AI tool updates enterprise courses and documents across languages with human review, audit trails and selective re-...
By Renata Fuchs 1 week ago
GPT-5.6 Sol tops Opus 5 on ARC-AGI-3 in OpenAI test
OpenAI says GPT-5.6 Sol reached 38.3% on ARC-AGI-3 using its Responses API, but the result depends on nonstandard settings.
By Colin Brandt 1 week agoWaymo AI evals drive release readiness, engineering leader says
Waymo’s Manasi Joshi said AI projects are judged by test maturity, with continuous evals, curated data and human release reviews.
By Colin Brandt 1 week ago
Google Lyria 3.5 music model adds section-level track edits
Google says Lyria 3.5 can generate longer, more controllable music tracks, but it has not detailed the model’s training data.
By Colin Brandt 1 week ago
PwC fabricated sources allegations hit Middle East AI reports
GPTZero says four PwC Middle East reports contained false or fabricated citations, including one it rated 84% likely AI-generated.
By Colin Brandt 1 week ago
Pangram AI text detector update claims one false flag per 24,000 documents
Pangram 4 costs $0.05 per 100 words and claims 99.66% AI-text detection, with Pangram 3 set to shut down Sept. 30, 2026.
By Wei-Lin Zhao 1 week ago
OpenAI Hugging Face breach involved credentials on four other services
OpenAI says internal AI agents found exposed credentials on four services during a security test, while Hugging Face traced 17,600...
By Wei-Lin Zhao 1 week agoTarget AI agents rely on governance, not just models, SVP says
Target SVP Siobhán Mc Feeney said AI advantage comes from agent governance, architecture and observability, with autonomy earned o...
By Colin Brandt 1 week ago
DeepMind AlphaFold team is broken up as authors leave for Anthropic
Google DeepMind has reassigned most AlphaFold paper authors, while John Jumper and colleagues move to Anthropic.
By Colin Brandt 1 week ago